g2n-native 0.1.0

Rust acceleration core for g2n (quantum statevector kernels, token sampling). Optional: g2n falls back to pure torch without it.

g2n-native

Rust acceleration core for g2n: in-place quantum statevector kernels and fast token sampling.

You normally don't install this directly — pip install g2n[native] pulls it in on supported platforms (Linux x86_64). g2n detects it at import time and falls back to its pure-torch paths if it's missing or G2N_NATIVE=0 is set.

API

• apply_gates(psi, gates) — apply a gate list in place to a (B, 2**n) complex64 C-contiguous statevector. Each gate is (bit_positions, matrix) with bit = n-1-qubit; matrices are (2,2)/(4,4) shared or (B,2,2)/(B,4,4) per batch row.
• sample_topp(logits, temperature, top_p, top_k=0, seed=0) — sample one token id from a (V,) f32 logits vector. temperature <= 0 means argmax. Deterministic for a given seed.

Parity with the torch reference paths is enforced by tests/test_native_parity.py in the g2n repo.

Building

maturin build --release --zig --compatibility manylinux_2_28

Proprietary — © g2n.dev.